MLOps Live (neptune.ai) · 2023
Jason Flaks on MLOps Live (neptune.ai): Deploying Conversational AI Products to Production
Jason Flaks, Co-Founder & CTO, appeared on MLOps Live (neptune.ai) in 2023 — a verified placement booked by Podlink for Xembly.
Jason Flaks, co-founder and CTO of Xembly, appeared on neptune.ai's MLOps Live for an interactive Q&A about deploying conversational AI products to production, published as a written article on September 6, 2023. Hosts Sabine and Stephen interview Flaks about Xembly, described as an automated chief of staff that handles tasks like meeting scheduling and automated note-taking. Flaks traces his background from music composition and math through audio signal processing, including work on the Xbox Kinect camera. He walks through Xembly's ML pipeline — speech recognition, speaker segmentation, formatting, then parallel paths for action-item extraction and abstractive recap summarization — and the infrastructure behind it, including Kubernetes, Kafka and AWS autoscaling. He discusses when to build in-house models versus using off-the-shelf systems, arguing business speech is a sub-domain where general models often mispredict, while fine-tuned large language models are the sensible choice for summarization. He also covers data labeling with Label Studio, testing live with silent-mode model deployments, and cautions about ChatGPT's convincing fabrications.
Listen to the episode
Key takeaways
- Xembly aims to give every worker an automated executive assistant, with conversational interaction and automated meeting notes as its two main product areas.
- Conversational AI is an unbounded domain, so fast iterative data labeling and multi-model pipelines are critical; Xembly keeps models independently upgradable.
- On build vs. buy: general-purpose ASR models often mispredict business speech, but Flaks says you'd "be silly" not to use a fine-tuned LLM for summarization.
- Testing is largely live: silent-mode parallel model runs, partial-traffic rollouts, and heuristics such as how often customers share meeting notes.
- Tools named include Kubernetes, Kafka, Faust, MongoDB, AWS, Kaldi, spaCy and Label Studio.
Want appearances like this?
This is what the booking service produces: real episodes on real shows, with clips that carry each interview past the show’s own audience.
Verified 2026-10-19 — every link above goes to the public episode.